Evozyne
Evozyne is a Chicago biotechnology company, founded by Paragon Biosciences in 2020, that designs new therapeutic proteins with generative AI and its own laboratory. EvoGen, its design engine, generates candidate proteins and estimates their potency, specificity, stability and immunogenicity, and EvoLab runs high throughput experiments whose results feed the next design round. Evo TestBench compares modeling approaches against experimental results across several kinds of design problem before a project commits to one. Partners gain access to the platform, its immunology expertise and its experimental capacity through research collaborations, which the company continues to seek in immunology, protein engineering and newer biologics.
Takeda's research and license collaboration on rare disease targets produced two highly functional proteins within six months. With NVIDIA, Evozyne built ProT-VAE, a transformer variational autoencoder on the BioNeMo framework, and its October 2025 PNAS paper reports a phenylalanine hydroxylase design with 2.5 times the catalytic activity of the natural enzyme and a carbonic anhydrase 61 °C more heat stable than the most stable known sequence. Gates Foundation supported work on heat tolerant proteins appeared in microPublication Biology.
Its own pipeline targets immune mediated diseases where biologics fail, including IgG mediated disorders, inflammatory bowel disease and engineered cytokines, with a first investigational new drug filing planned for 2026. Marc Yoskowitz is chief executive, and an $81 million Series B in 2024 was led by OrbiMed and Fidelity with NVIDIA's NVentures taking part.
Capability Axes
An AI Health Index grade measures what a buyer can verify from public sources on the date shown. It is not a rating of how good the product is. A vendor can build an excellent system and grade low on an axis because it publishes nothing an outsider can check. How grades read
Generative protein design is what a partner buys into. EvoGen generates candidate proteins for a therapeutic objective and estimates their potency, specificity, stability and immunogenicity so the team can choose which designs to make, and EvoLab's high throughput experiments test those designs and feed the results into the next round.
A collaboration delivers tested research candidates rather than a model subscription, with experimental capacity and scientific expertise included, but turning a research objective into a set of proposed proteins in the first place is the model's work.
Model proposals pass through laboratory evaluation before a project judges whether a candidate has the intended properties: EvoLab runs the tests, and research teams use the measurements to pick which designs to pursue. Evo TestBench compares model performance against experiment so researchers choose an approach suited to the project, and scientific decisions and the move into development stay with the research team. No standard partner approval threshold or quantitative stop rule decides which proposals enter a new laboratory stage; each project sets its own gates.
EvoGen is the generative design engine, and the ProT-VAE work with NVIDIA, published in PNAS, names its transformer and variational autoencoder architecture built on the BioNeMo framework. Experimental results feed back to refine the models, and Evo TestBench compares several approaches instead of treating one as right for every project, so the technology behind a collaboration varies with the task. There is no full list of production models, release history or account of the datasets behind a partnership, so a partner cannot tie a change in output to a model update.
Evozyne runs EvoGen inside its own research and laboratory operation, an in house chain from model to experiment. NVIDIA helped develop the ProT-VAE architecture and its BioNeMo framework underpins the models, as the published research sets out, which identifies a development partner rather than a party receiving current partner inputs. No full list of compute hosts, model components and data processors reaches a partner, and no standard term names outside recipients or their retention and training rights for each step.
The Takeda research and license collaboration is named, and Evozyne reports it produced two highly functional proteins within six months. With NVIDIA, the company published ProT-VAE, the architecture behind EvoGen, in PNAS in October 2025: a designed phenylalanine hydroxylase reached 2.5 times the catalytic activity of the natural enzyme, and a carbonic anhydrase gained 61 °C in melting temperature over the most stable known sequence.
Gates Foundation supported work appeared in microPublication Biology. These are laboratory results for parts of the platform, with no evidence yet for the current workflow's clinical candidates or for benefit in patients.
EvoLab returns experimental observations on research candidates to the discovery program, and those measurements inform the next EvoGen rounds. The privacy policy treats that experimental data separately from information visitors submit through the website, and its retention language reaches only the website.
No standard collaboration term sets de identification for contributed patient information, a retention period for partner research inputs or limits on reusing partner datasets, and the platform's safety testing concerns candidate properties, with no stated process for a partner data incident.
Pharmaceutical and biotechnology partners reach the design platform through research collaborations that test candidate proteins and compare their experimental properties, with no clinical record workflow or patient portal. The general privacy policy covers visitors, contacts and site users and sets no terms for protected health information in a partner project. No business associate relationship or BAA tier is stated, and no collaboration commitment says which identifiable patient information, if any, a partner may introduce.
A collaboration can carry commercially sensitive target information, research inputs and experimental results through a scientific workflow Evozyne runs on its own proprietary datasets. The website privacy policy covers contact and browsing data, and research terms depend on each partner agreement.
No SOC 2 report, ISO 27001 certificate or security audit covers EvoGen and EvoLab, and there is no stated cadence of independent security testing or description of the controls on collaboration data moving between modeling and the lab.
EvoGen and EvoLab support discovery and preclinical testing, and partner work concerns research candidates for later development. Evozyne plans its first investigational new drug submission in 2026 for an internal immunology program, a filing that would cover a drug candidate rather than the research platform. Selecting a candidate says nothing yet about its safety or effectiveness in patients, and the platform carries no diagnostic or treatment device indication.
Evo TestBench is a defined process for comparing modeling approaches against experimental results across several categories of therapeutic design problem, and it decides which models a project relies on, since approaches perform differently from task to task. EvoGen also carries immunogenicity prediction modules, with stability and sequence risks tested early and often. The program is internal, with no external audit, and its testing covers laboratory design tasks rather than how a candidate behaves across the patient populations a future treatment would reach.
Experimental testing lets the research team reject an AI proposal that misses the project's scientific objective and send a candidate through another round. In a partnered program such as Takeda's, rights and obligations sit in a negotiated research and license agreement, while the website disclaimer reaches only site content and use. No accuracy guarantee or stated remedy covers a partner whose candidate fails a design or laboratory objective, and the terms of any commissioned program stay inside its own contract.
The customer workflow is a research collaboration, with software and laboratory work running inside Evozyne's own discovery process and results evaluated with the partner. Nothing in it works from a patient chart: there are no EHR connections, HL7 or FHIR interfaces or exchange with a health system's clinical software, and laboratory data moves back into modeling internally. Partners receive research results, not a tool installed in clinical systems.
Collaborations run on a discovery environment and laboratory Evozyne operates in Chicago, with partners working through the project team and results feeding later EvoGen rounds. EvoGen's model was built on NVIDIA's BioNeMo framework, which says nothing about where partner data is stored or processed. There is no customer hosted or local version, no committed storage or inference region and no stated method for isolating one partner's computation and research material from another's.
Partners commission research through a collaboration that can draw on EvoGen, EvoLab and Evozyne's scientists, with access and intellectual property governed by a negotiated license, and the Takeda agreement was built around a defined group of rare disease targets. The company seeks collaborations in immunology, protein engineering and newer biologics, and a buyer reaches the team directly rather than opening a software account. No standard platform license fee, per project price or milestone schedule is stated for a new collaboration.
The setting is biopharma research and preclinical development. Internal programs focus on immune mediated diseases where biologics fail or lose effect, including IgG mediated disorders, inflammatory bowel disease and engineered cytokines, while the Takeda collaboration applied the approach to rare disease targets and the NVIDIA work to enzyme design in the lab. Partners without a comparable platform get scientific and laboratory support with the research. Results so far cover parts of discovery, with no clinical validation across the diseases on the pipeline.
Alternatives to Evozyne
Other products that do the same kind of work as Evozyne, and what each one does.
Scala Biodesign makes software that redesigns proteins to work better, and sells it both as a cloud platform and as a design service.
BioGeometry builds generative AI models for protein design and uses them in partner projects and its own drug pipeline.
Earendil Labs uses AI to design antibody drugs and develops them through its own pipeline, licensing individual programs to large pharmaceutical companies.
Terray Therapeutics is a Los Angeles small molecule company whose founding argument is about data rather than models. Its chief executive Jacob Berlin put it plainly at the Series A: even the most powerful artificial intelligence cannot get past insufficient or unclear data.
Variational AI develops Enki, a generative model for designing small molecules against a research team's target product profile.
immunitoAI designs antibody drugs with AI and tests the designs in its own laboratory.
Pricing
Vendor-published figures are labeled as such. Figures labeled “Estimated” are derived from third-party sources and have not been confirmed by the vendor.
| Entry Price | Pricing Basis | BAA Tier | Implementation | Source |
|---|---|---|---|---|
|
Negotiated research collaboration
|
Partners negotiate the scope of a discovery program and any license rights. The Takeda agreement was organized around a defined group of rare disease targets. A standard rate is unavailable. | The offering is a research collaboration with no stated BAA tier. | EvoGen modeling and EvoLab experimental work are part of the negotiated program. Separate onboarding or laboratory fees depend on the agreement. | Vendor Published |
Evozyne supplies research and laboratory capacity through collaborations. Buyers engage directly to define scope, deliverables and licensing. There is no standard software seat price.